A novel automatic approach for glioma segmentation

Abstract

The quantitative analysis of brain magnetic resonance imaging (MRI) represents a tiring routine and enormously on accurate segmentation of some brain regions. Gliomas represent the most common and aggressive brain tumors. In their highest grade, it can lead to a very short life. The treatment planning is decided after the analysis of MRI data to assess tumors. This treatment is manually performed which needs time and represents a tedious task. Automatic and accurate segmentation technique becomes a challenging problem since these tumors can take a variety of sizes, contrast, and shape. For these reasons, we are motivated to suggest a new segmentation approach using deep learning. A new segmentation scheme is suggested using Convolutional Neural Networks (CNN). The presented scheme is tested using recent datasets (BraTS 2017, 2018, and 2020). It achieves good performances compared to new methods, with Dice scores of 0.86 for the Whole Tumor, 0.82 for Tumor Core, and 0.6 for Enhancing Tumor based on the first dataset. According to the second dataset, the three regions had an average of 0.88, 0.77, and 0.65, respectively. The new dataset provides 0.87, 0.91, and 0.79 for the three regions, respectively.